The Best AI Agent Frameworks

Large Language Models (LLMs) are an amazing technology, but a plain model is just like a brain in a jar, unable to interact with the environment around it. AI agents are the next step — they’re autonomous systems using LLMs to understand their environment, make decisions, plan actions and call available tools to achieve user-defined goals.

Tools are just any functions with a clear description of how to use them. The AI agent can call one of its tools when it decides it’s necessary to accomplish the goal — for example, when asked by the user for ideas on what to wear today, it can retrieve current weather data to make an appropriate decision. Tools are the body that we give to the brain — the LLM.

AI agents are already present in various use cases, including intelligent assistants, web browsing, task management, code generation, software design, personalized marketing, streamlining repetitive tasks and many more.

How AI agents work

AI agents use a few key concepts to reach a user-defined objective. Generally, when faced with a task, the agent starts by reasoning and planning on how to achieve the objective. It plans the steps it’s going to take and tools it’s going to use from the available set. For instance, in order to give you outfit ideas, it can come up with a plan like that:

1. Call the weather-checking tool to retrieve information about the weather
2. Call the web search tool to retrieve information about current fashion trends
3. Compile an outfit
4. Return the answer to the user

Steps 1 and 2 perform tool-calling and the agent obtains results from the tools.

Since responses from both step 1 and 2 are required to perform step 3, it also needs memory. Most often, memory is short-term and features all the previous messages from one run of the agent, but you can also have long-term memory which preserves some information across conversations, like your fashion taste, or you can manipulate the messages you pass to the agent at a time to use fewer tokens and produce more precise results, especially when not all information is needed at once.

Another common feature is retrieval — you can use Retrieval-Augmented Generation technique to provide the agent with some custom knowledge. It modifies the query to the LLM to contain relevant context from available documents.

AI agentic frameworks

For simple use cases, predefined agent workflows — the things the agent does and their order — are enough. For simpler use cases, like a chain of prompts, you can even use plain code. But when the workflow becomes more complex, you should use an agentic framework which gives you more control and streamlines development. Frameworks offer tool parsing, easy access to metadata like token usage, structuring output based on your needs, easier multi-agent cooperation, conditional execution and more.

Choosing the right AI agent framework

AI agent frameworks come in all shapes and sizes — choosing the right one may prove crucial to your use case. There are a few key points to consider:

Ease of use

AI frameworks range from ones where you don’t even need to know how to code, through ones where very little coding skills are required, to ones where more in-depth knowledge of typing, objective-oriented programming and basic proficiency in graphs is needed.

Task complexity

While simpler frameworks may be enough for a basic use case, they aren’t designed for all tasks and you may benefit from more in-depth design of the agent. When a decision must be made, a loop executed, or a check performed, it’s generally better to enforce it in the workflow architecture rather than relying on the agent to devise it each time. Basically, well-known tasks should use a predefined flow rather than a freshly generated one. If you don’t know which tasks will be repetitive in advance, you can start with a simpler framework, observe which activities repeat, and then hardcode them into the workflow using a more advanced AI agent framework — this is how we approached our first agents at Software Mansion.

Community

Sometimes documentation isn’t enough, and you may benefit greatly from an active community of framework users that have already faced many of the same challenges you may face. It’s also super helpful to have a space where you can ask questions someone may know the answers to. Active community also offers smaller probability of bugs and vulnerabilities in the framework code.

Performance

Approaches offered by AI frameworks vary in latency, that is the time the user has to wait for completion. As a rule of thumb, agentic frameworks where you have more control lead to lower latency, while frameworks that are more autonomous tend to perform some redundant actions that increase it. Precisely designed workflows can also offer results that are more satisfactory to the user and stable across agent invocations, as they follow a stricter routine.

Token usage

Simpler frameworks tend to use up more tokens, as their memory is less selective and agents without a hardcoded workflow tend to make mistakes and repeat some steps.

Scalability

Increased token usage and higher latency are serious obstacles when it comes to scalability — if you want to create something more than a personal agent for a small task, using more advanced AI agent frameworks is recommended. Task complexity also usually increases with scale, which is really prone to generating redundant repetitions in case of too autonomous agents.

Integration

Various frameworks offer different levels of integration with other technologies — the more complex your project is, the more important integration is.

Privacy and security

Frameworks also differ in their handling of sensitive information, encryption of data both at rest and in transit or access control. This may be an issue depending on the sensitivity of the data you use in your task. Giving an agent too much freedom when handling sensitive data may even cause it to inadvertently ‘spill’ that data.

The best frameworks for AI agents

The following list presents the best frameworks you can use to build AI agents, but the order doesn’t represent increasing or decreasing overall judgment of the framework, but rather increasing complexity — more control over the workflow and more functionalities. The best choice for you depends on the points explained above, and this list is meant to help you decide based on your needs and skills.

RelevanceAI

RelevanceAI is the choice to make when you’re a non-technical person. It doesn’t require any code to create agents and instead comes with an UI for agent & tool creation. It offers fast deployment and is pretty business and user-friendly. However, it’s not a good choice for more complex systems and doesn’t have a huge community of developers. If you’re concerned about low latency or minimal overhead, this framework may not be for you. RelevanceAI isn’t open source, which may make it harder to detect bugs in the framework. You can learn about RelevanceAI here.

smolagents

Smolagents is a minimalist AI agent framework by Hugging Face. It offers both typical tool-calling and code agents, which can generate and execute code where they may also use the provided tools. It comes with a really good tutorial and introduction to the agents themselves. When you’re a developer or a team of developers wishing to start building agents, smolagents is definitely the point to start — you can learn a lot about the way agents operate, see if they’re a good solution for your problem and later switch to more advanced frameworks. Its agents are very autonomous, so we generally recommend using it for simple use cases or as a learning tool. Hugging Face backing is a great sign of credibility.

PhiData

PhiData enables you to build AI assistants with memory, tools, knowledge, reasoning or multi-agent orchestration. It’s designed to make it simpler to turn LLMs into assistants. It makes handling of memory, tools and UI easier than coding it yourself. The community is emerging and the framework is open-source.

LangChain

Currently the most widely-adopted agentic framework. It allows developers to chain prompts, models, memory and tools. It offers native support for most popular LLM providers and comes with a rich ecosystem of tools agents may use. It’s very modular and is a good solution for medium-complexity tasks. You need to understand a few abstractions like chains, agents, prompt design, memory, retrievers etc., but the community around it is already mature and there are lots of tutorials. Integration and extensibility are also the strong suits of LangChain. You can learn about LangChain here.

LlamaIndex

LlamaIndex is heavily focused on data ingestion, indexing, retrieval, query engines and document handling. Building RAG applications with it is pretty easy, but it might be less suited as an orchestration framework for many agents with distinct roles. Performance depends on the quality of documents provided and embedding models. It allows extending with your custom indexing, retrievers etc. The community is strong, but the framework may not be the best solution for branched or highly-orchestrated workflows. You can learn about LlamaIndex here.

CrewAI

CrewAI is designed for high-level multi-agent orchestration with roles/crews, where LLMs are used to coordinate the work of agents. It’s newer than LangChain, but popular nevertheless. You can build multi-agentic AI systems pretty fast with it, but with very large tasks the performance may still be suboptimal, as it doesn’t offer flow control and agents self-organize everything. It doesn’t offer parallel execution. You can learn more about CrewAI here.

LangGraph

LangGraph is an extension of LangChain and the most advanced AI framework on the list — it allows you to define agentic workflows as graphs, where each node represents a step in execution and edges define possible transitions. Agents in LangGraph are stateful — apart from history of messages, you may track anything else you want, like the number of iterations some loop in the graph has been invoked. LangGraph offers conditional edges, loops, persistent state, easy debugging, custom tools, structured output and much, much more. It’s the technology we use at Software Mansion to build our agents. LangGraph, along with LangChain as the base and LangSmith as the tool to easily observe in-depth execution, constitutes a complete environment for building advanced agents. The highest level of control makes it great for scalable solutions where minimal latency is also crucial.

Why are we the right people to talk about top AI agent frameworks? At Software Mansion, we’ve already built a few innovative projects with AI. We created React Native ExecuTorch, which lets you run AI models and LLMs on-device in React Native with ExecuTorch. We also built React Native RAG to improve LLM capabilities with on-device Retrieval-Augmented Generation, and Private Mind, a chat app that runs LLMs entirely locally on your device.

We are Software Mansion — software development consultants, a team of React Native core contributors, multimedia and AI experts. Drop us a line at [email protected] and let’s find out how we can help you with your project.